Enough users will end up using certain images such as the tiger making it worthwhile to fake a prompt.
"The user is then presented 10 images (a tiger,
a house, a moose, etc) from a library of 10,000 images."
I should have made this more clear. The ten images are chosen randomly from the group of 10,000The question is: are there 10,000 images that are different enough people won't be fooled. Say my picture is a green house. And a prompt has a picture of a red house, will I accidentally think its the right site key?
The good news is most people won't even have a picture of a house as their site key so it will protect a large percent.
1) If the attacker can scrape the screen, they can detect which image you are using - securing the entire pipeline to the screen is hard.
2) 10,000 images is way too few.
Even if we can assume an even distribution of images, as an attacker I can serve the same image to all targets, 1 in 10,000 will now think that they are interacting with a trusted component